Healthcare and Artificial Intelligence

By Cédric Villani and Daniela Rus

This book provides an overview of the role of AI in medicine and, more generally, of issues at the intersection of mathematics, informatics, and medicine. It is intended for AI experts, offering them a valuable retrospective and a global vision for the future, as well as for non-experts who are curious about this timely and important subject. Its goal is to provide clear, objective, and reasonable information on the issues covered, avoiding any fantasies that the topic “AI” might evoke. In addition, the book seeks to provide a broad kaleidoscopic perspective, rather than deep technical details.

 

• Language: English
• Publisher: Springer
• Release date: Mar 17, 2020
• ISBN: 9783030321611

Book preview

Editors

Bernard Nordlinger, Cédric Villani and Daniela Rus

Healthcare and Artificial Intelligence

../images/485811_1_En_BookFrontmatter_Figa_HTML.png

Editors

Bernard Nordlinger

Department of Surgery, Hôpital Ambroise Paré, Assistance Publique Hôpitaux de Paris, Boulogne-Billancourt, France

Cédric Villani

Institut Henri Poincaré, Université de Lyon, Paris, France

Daniela Rus

Computer Science and Artificial Intelligence, Massachusetts Institute of Technology, Cambridge, MA, USA

ISBN 978-3-030-32160-4e-ISBN 978-3-030-32161-1

https://doi.org/10.1007/978-3-030-32161-1

© Springer Nature Switzerland AG 2020

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed.

The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use.

The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG

The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Foreword

This book is the result of a meeting between the mathematical Fields Medalist Cédric Villani, the brilliant surgeon Bernard Nordlinger, and the MacArthur Genius award-winning roboticist and computer scientist Daniela Rus. A priori, there was little to suggest they would ever work together: Cédric Villani was promoting mathematics as the head of the Raymond Poincaré Institute, Bernard Nordlinger was trying to improve the prognosis of cancers through ambitious therapeutic trials, and Daniela Rus was trying to create widely available robots and AI systems to support people with physical and cognitive work.

However, their meeting was not purely fortuitous since their trajectories crossed at an event in support of innovation and research. Bernard Nordlinger whose concerns extended beyond surgical procedure had extensive experience in controlled therapeutic trials. He was well aware of methodological requirements and the most appropriate means of exploiting the results. Cédric Villani whose interests went well beyond the world of mathematics was seeking advances that had the potential to be enabled across many fields (particularly, health). Daniela Rus was looking for advances having a societal impact (particularly, addressing significant challenges around curing disease).

Although their complementarity was obvious, it was necessary to provide a means of establishing exchanges between those who accumulated data and those who actively worked in particular fields. Cédric Villani and Bernard Nordlinger set up a joint working group between the French Academy of Sciences and the French Academy of Medicine. The working group comprised a number of highly qualified members of the two academies and many other experts. The objective of the working group was to meet to gather information, listen to presentations, and exchange ideas. Daniela Rus joined later and brought an American viewpoint and expertise.

A joint meeting of the two academies on Mathematics, Big Data, and Health: The Example of Cancer took place on November 28, 2017. By presenting examples from health insurance databases pertaining to results already obtained in imaging and genetics, they provided us with the first opportunity to highlight the prospects that mathematics could offer to the progress of science and technology in health.

The characteristics of health sciences and their structural wounds as Cédric Villani put it (more particularly, the complexity of phenomena, the variability of conclusions, the complexity of funding, and human interpretative bias) must find a new dynamic of progress by mobilizing mathematical sciences. By making it possible to exploit Big Data and, more generally, artificial intelligence, medicine will increase its means of action. No field of activity will be able to escape this evolution. The very special resonance that artificial intelligence will have in the health field should be obvious to all.

Renewing and perfecting the interpretation of images; increasing performance in radiology, pathological anatomy, and dermatology; taking advantage of genetic data; and developing precision medicine all will become possible thanks to artificial intelligence. It will enable the collection of data of a previously inaccessible richness to make possible the most careful clinical examinations. It will make an irreplaceable contribution to diagnostic and therapeutic choices.

Such a development presupposes the retraining of doctors and, more broadly, of health professionals. Such a change in practices also requires consideration of the evolution of responsibilities and their legal consequences. If the robot and the algorithm acquire such a critical role in the decision, will they assume a legal personality? The hopes raised by artificial intelligence raise questions and even concerns. These concerns are justified. The basic precaution is to be careful as we move toward the fantastic predictions of the augmented man that on the basis of adventurous extrapolations suggest a totally uncontrolled evolution.

We must thank Bernard Nordlinger, Cédric Villani, and Daniela Rus for getting some of the top experts to participate and gathering a series of solidly argued clear articles that should allow readers to learn calmly without reserve and without passion about the prospects for the development of artificial intelligence in health.

Let us hope that this book will convince its readers that there is no question of replacing the doctor with the machine and that the challenge is to organize the natural and collaborative interactions between human expertise and the contributions of artificial intelligence in the daily practice of medicine. ¹

Daniel Couturier

Catherine Bréchignac

Paris, France

Introduction

Bernard Nordlinger

Cédric Villani

Daniela Rus

Artificial intelligence (AI) is a revolution for some people, fashion for others, and a reality for many aspects of our lives. All areas scientific and otherwise are preparing to experience the upheavals it will inevitably bring.

At the forefront of scientific fields awaiting this impact is health. On the one hand, there are the usual problems such as diagnostic errors, variability of situations, fallibility of experts, and great difficulties in transmitting research information to practitioners. On the other hand, AI excels at digesting piles of literature, finding rare correlations, and analyzing images and other data that are ever more numerous produced by medicine, a field in which the stakes are of course literally vital.

This book provides an overview of AI in medicine and, more generally, looks at issues at the intersection of mathematics, informatics, and medicine. It reaches out to AI experts by offering hindsight and a global vision, and to non-experts intrigued by this timely and important subject. It provides clear, objective, and reasonable information on issues at this intersection avoiding any fantasies that the AI topic may evoke. The book provides a broad kaleidoscopic viewpoint rather than deep technical details.

The book was gradually compiled by the Artificial Intelligence and Health working group created by the National Academy of Medicine and the Academy of Sciences of France at our suggestion. The sessions of the working group were an opportunity for us to explore the questions we had asked ourselves during our meeting at the Ethics Council of the Epidemium competition. Above all, they were a means of inviting recognized specialists to contribute and collaborate such as mathematicians, modelers and analysts, data scientists, statisticians, oncologists, surgeons, oncogeneticists, sociologists, hospital administrators, and lawyers.

For the applications of AI to be impactful a multidisciplinary approach is required. This is the reason we set up a multidisciplinary working group with specialists from various fields of medicine, mathematics, and computation. Our conversations were fruitful because we were able to discuss topics that are intrinsically multidisciplinary with experts representing all their medical and computational aspects. The dialogue of this think tank made us realize that the doctors of the future will have to be multilingual in medicine, data, and computation. This will require new approaches to training medical students, something we believe should be done early in medical training.

The road ahead is full of challenges, but the journey is worth it. The mechanical medical doctor (MD) is certainly not for tomorrow and certainly not desirable. The doctor of the future should be augmented, better equipped, and well informed to prevent, analyze, decide, and treat disease with empathy and the human touch. The aim will be to improve diagnoses, observations, therapeutic choices, and outcomes.

Imaging and, more generally, the analysis of medical data will benefit from AI and machine learning advances. Medicine is an area of choice for the full exploitation of images because medical image details are too fine or subtle to be picked up by the naked eye. Additionally, when image resolution is poor, algorithms can identify the missing data and enhance image quality. Mathematical models can then be used to predict future outcomes given the current images.

Modeling can be used as an enhanced tool for the evaluation of chemotherapy treatments. Right now tumor response to chemotherapy is  estimated using approximate measurements of the diameter of the tumor. New advances in tumor growth modeling processes make it possible to predict tumor response to treatments that take into account macroscopic features. Equally important is the consideration of tumor genomic data or modeling. In the future these predictions will take into account information at both the molecular and the macroscopic scale and will be a significant challenge.

The correspondence between genotype and phenotype remains the most vexing of the mysteries of biology, but new statistical learning methods offer a workaround. They enable the analysis of extraordinary molecular datasets to be self-mathematized to better identify the regions of the genome associated with tumor progression and determine the most appropriate treatment. This same machine learning approach can help to overcome the imbalance between the large number of variables studied and the small sample sizes available by selecting the most relevant data. The future of clinical research will be enabled by tumor databanks. Combining clinical and biological data stored in tumor banks will likely become standard practice for clinical research. Clinical trials are in the process of being renewed: the platforms that make it possible to record and cross-reference large clinical and molecular data suggest new trial formats. Studies target molecular abnormalities more than organs. When strong pilot alterations are identified, molecular information will gradually enter into common practice.

Pathology will also be renewed by AI. Today pathologists depend on physical slides that cannot be shared. Imaging the slides will in future lead to pathology databanks. Such virtual slides will allow remote analysis without a microscope. Multiparametric analysis makes it possible to search for increasingly complex information. New technologies for organ exploration that are extremely promising are beginning to emerge yielding large amounts of data, in addition. The microbiota, the environment, and now entire ecosystems can be identified by statistical tools.

Finally, and perhaps most importantly, medical data are becoming accessible, especially to young multidisciplinary teams working in open mode (open science, open data).

Such developments will have to face up to major challenges associated with the race for the three major ingredients that are the object of unbridled competition worldwide: human expert brains, high-performance computer equipment, and very large multiparameter datasets.

First of all is the challenge of finding human expert brains. AI is mainly carried out by experts endowed with a spirit of collaboration, ready to take on a challenge, and with expertise in interdisciplinary contacts. Dual-competency profiles for new jobs (health and Big Data, algorithms and medicine, etc.) are very popular, yet the talent pool is scant. Training a new generation of scientists ready to occupy these interfaces is a major challenge. There are not enough trainers or programs; we will have to start new experimental training programs.

Next is the challenge of producing high-performance computer equipment. Europe lags behind the United States and China in terms of investment. The AI of health in the future will require large storage and computing centers, some of which will be internal and some will be outsourced.

Last but not least is the search for machine learning algorithms to identify correlations, regularities, and weak signals. This will require very large multiparameter datasets, labeled data, and cross-referencing the data of various sources (private and public) whose marriage can only be accomplished with governmental support. In France the largest existing public health databases are interlinked such as SNIIRAM (National Information System Inter Plans Health Insurance), SNDS (National Health Data System), and hospital platforms such as the platform of the AP-HP (Administration of Paris Public Hospitals). They are all intended to be part of a larger entity aimed at facilitating their access for research purposes.

The construction of such large medical datasets presents a triple challenge. First, the technical and technological challenge posed by the size of the data has to be resolved. This includes rapid access to data, handling a variety of formats, addressing the considerable number of imperfections in the data, and the need for enhanced cybersecurity. Second, the ethical and legal challenge relating to the protection of personal data requires solutions that ensure public confidence. Third, trust in sharing data is difficult to come by in practice since it requires convincing data owners to overcome their reluctance to join forces. In the French context the State is playing a key role in all these areas. France has set up the Health Data Hub to facilitate research.

Finally, patients will have to play their part in such developments. This includes giving informed consent, participating in the development of tools, and contributing to the evolution of practices and mentalities. While physicians retain responsibility for contact with and explanations to patients, an interaction between humans and algorithms in medicine that is appropriate and constructive is likely to emerge (especially defined for the human dimension).

Today, all nations aspiring to be part of the global innovation scene are embarking on AI and health programs. The United States and China have announced national strategies to stimulate the development and applications of AI with medicine and healthcare as important pillars. France is no exception as shown by releasing the report Giving meaning to artificial intelligence (by Cédric Villani) in March 2018. The findings in the report are the result of six months of interdisciplinary work including many hearings. This report outlines a strategy for a coordinated AI policy in France and, more broadly, in Europe. The President of the French Republic adopted most of the recommendations in his speech on March 29, 2018.

The multidisciplinary working group responsible for the genesis of this book contributed some of the background and reflections that shaped the report sent to the government; it also influenced details of the Data Protection Act that adapts French law for the purposes of the EU General Data Protection Regulation (RGPD). It is intended to provoke advance thinking of both specialists and the merely curious. This book brings together a variety of very brief contributions proposed by stakeholders to the working group. All the themes already outlined are included.

The last part of this book examines breakthroughs that connect medicine and society such as the ambivalent role of new modes of information (particularly, social networks that disrupt communication and spread fake news and conspiracy theories), the emotional and sometimes misleading dialogue between humans and machines, and humanity’s temptation to push ahead without safeguards. Throughout the book, contributors emphasize directions and dominant themes more than details.

The purpose of this book is to give the reader an overview of the state of the art of AI in medicine by providing examples and suggestions of how the medical field will change. Machines supporting and augmenting doctors should not be seen as scary. Rather, the future augmented doctor will provide patients with better and more personalized treatments.

Contents

Artificial Intelligence and Tomorrow’s Health 1

Cédric Villani and Bertrand Rondepierre

Advancing Healthcare Through Data-Driven Medicine and Artificial Intelligence 9

Ran D. Balicer and Chandra Cohen-Stavi

Artificial Intelligence:​ A Vector for Positive Change in Medicine 17

Daniela Rus

Databases

Machine Learning and Massive Health Data 23

Emmanuel Bacry and Stéphane Gaïffas

Linking an Epidemiological Cohort to Medicoadministra​tive Databases 33

Marie Zins and Marcel Goldberg

Medical and Administrative Data on Health Insurance 43

Claude Gissot

How the SNIIRAM–PMSI Database Enables the Study of Surgical Practices 51

Bertrand Lukacs

Hospital Databases 57

Christel Daniel and Elisa Salamanca

Experience of the Institut Curie 69

Alain Livartowski

Knowledge by Design 75

Jean-Yves Robin

Diagnosis and Treatment Assistance

Artificial Intelligence to Help the Practitioner Choose the Right Treatment:​ Watson for Oncology 81

Christine Garcia and Georges Uzbelger

Artificial Intelligence to Help Choose Treatment and Diagnosis Commentary 85

Jacques Rouessé and Alain Livartowski

Medical Imaging in the Age of Artificial Intelligence 89

Nicholas Ayache

Artificial Intelligence in Medical Imaging 93

Johan Brag

The AI Guardian for Surgery 105

Daniela Rus and Ozanan R. Meireles

Toward an Operating Room Control Tower?​ 115

Nicolas Padoy

High-Dimensional Statistical Learning and Its Application to Oncological Diagnosis by Radiomics 121

Charles Bouveyron

Mathematical Modeling of Tumors 129

Thierry Colin

Toward an Augmented Radiologist 141

Mostafa El Hajjam

Functional Imaging for Customization of Oncology Treatment 145

Dimitris Visvikis and Catherine Cheze Le Rest

Virtual Image and Pathology 153

Cécile Badoual

Research

Artificial Intelligence and Cancer Genomics 165

Jean-Philippe Vert

Besides the Genome, the Environmentome?​ 175

Alain-Jacques Valleron and Pierre Bougnères

Artificial Intelligence Applied to Oncology 185

Jean-Yves Blay, Jurgi Camblong and François Sigaux

Physical and Mathematical Modeling in Oncology:​ Examples 199

Martine Ben Amar

Artificial Intelligence for Medical and Pharmaceutical Research:​ Can Artificial Intelligence Help in the Discovery and Development of New Drugs?​ 207

Gilles Wainrib

Open Science and Open Data:​ Accelerating Scientific Research 215

Mehdi Benchoufi and Olivier de Fresnoye

Data

Data:​ Protection and Legal Conditions of Access and Processing 225

Jeanne Bossi-Malafosse

CNIL (Commission Nationale de l’Informatique et des Libertés) and Analysis of Big Data Projects in the Health Sector 235

Matthieu Grall

International Vision of Big Data 241

Laurent Degos

Human

Are Public Health Issues Endangered by Information?​ 255

Gérald Bronner

Social and Emotional Robots:​ Useful Artificial Intelligence in the Absence of Consciousness 261

Laurence Devillers

Augment Humans, and Then What?​ 269

Guy Vallancien

Artificial Intelligence:​ A Vector for the Evolution of Health Professions and the Organization of Hospitals 273

Jean-Patrick Lajonchère

Editors and Contributors

About the Editors

Bernard Nordlinger

Professor of Surgical Oncology, Université de Versailles, AP-HP; Coorganizer of the IA and Health Working Group of the Académie nationale de Médecine and Académie des sciences.

Cédric Villani

mathematician; Fields Medal winner; professor at the Claude Bernard Lyon 1 University; member of the Academy of Sciences; Member of the French Parliament; first Vice President of the Office parlementaire d’évaluation des choix scientifiques et technologiques (OPECST).

Daniela Rus

Director of the Computer Science and Artificial Intelligence Laboratory (CSAIL); Andrew and Erna Viterbi Professor in the Department of Electrical Engineering and Computer Science (EECS) at the Massachusetts Institute of Technology.

Contributors

Nicholas Ayache

Research Director, INRIA, Sophia Antipolis, Member of the Academy of Sciences. Scientific Director of the AI institute 3IA Côte d’Azur.

Emmanuel Bacry

Senior Researcher at CNRS, University of Paris-Dauphine, Chief Scientific Director Health Data Hub, Head of Health/Data projects, École Polytechnique.

Cécile Badoual

Department of Pathology, University of Paris Descartes, Georges Pompidou European Hospital, APHP.

Ran D. Balicer, M.D., Ph.D., MPH

Director, Clalit Research Institute, Director, Health Policy Planning Department, Clalit Health Services, Chief Physician office, Associate Professor, Epidemiology Depart- ment, Faculty of Health Sciences, Ben-Gurion University of the Negev.

Martine Ben Amar

Professor of Physics, Laboratory of Physics, École normale supérieure et Sorbonne University; Institute of Cancerology of Sorbonne University.

Mehdi Benchoufi

Faculty of Medicine, University of Paris Descartes, INSERM UMR1153, Centre d’Épidémiologie Clinique, Hôpital Hôtel Dieu, AP-HP, coordinator of the Epidemium program.

Jean-Yves Blay

Director General of the Léon Bérard Centre, Groupe Sarcome Français, NETSARC+, EURACAN, Centre de Recherche en Cancérologie de Lyon, LYRI- CAN, Université Claude Bernard, Lyon 1.

Jeanne Bossi-Malafosse

Lawyer specializing in personal data and health information systems, DELSOL Lawyers.

Pierre Bougnères

Pediatric Endocrinologist, Professor of Pediatrics at the University of Paris-Sud, Inserm Unit U1169 Gene Therapy, Genetics, Epigenetics in Neurology, Endocrinology and Child Development.

Charles Bouveyron

Professor of Applied Mathematics, Inria Chair in Data Science, Laboratoire J.A. Dieudonné, UMR CNRS 735, Equipe Epione, INRIA Sophia Antipolis, Université Côte d’Azur, Nice, France.

Johan Brag

Chief Scientific Officer, Median Technologies, Valbonne, France.

Gérald Bronner

Professor of Sociology, University of Paris 7, Académie des Technologies, Académie nationale de Médecine.

Jurgi Camblong

CEO and Co-founder of SOPHiA GENETICS, member of the Council for Digital Transformation, Swiss Federal Government.

Chandra Cohen-Stavi

Clalit Research Institute.

Thierry Colin

Senior Vice-President, Radiomics Business Area, SOPHiA GENETICS.

Christel Daniel, M.D., Ph.D.

Deputy Director of the Data and Digital Innovations Departement—Information Systems DirectionAP-HP, LIMICS INSERM UMRS 1142.

Laurent Degos

Academy of Medicine. Academy of Sciences (correspondent) Professor Emeritus University of Paris. Former President of the High Authority of Health.

Laurence Devillers

Professor of Artificial Intelligence and Ethics, Sorbonne-Université/LIMSI-CNRS, member of CERNA-Allistène, DATAIA mission member, founding member of HUB France IA, IEEE P7008 Nudging member.

Olivier de Fresnoye

Coordinator of the Epidemium programme. Ozanan Meireles, surgeon, MGH.

Mostafa El Hajjam

Interventional Radiologist, Department of IMAGERIE, Ambroise Paré Hospital, Boulogne.

Stephane Gaïffas

LPMA Professor, University of Paris-Diderot. Christine Garcia, IBM France, consultant and teacher, member of Syntec Numérique Santé.

Christine Garcia

IBM France, consultant and teacher, member of Syntec Numérique Santé.

Claude Gissot

CNAM, Director of Strategy, Studies and Statistics.

Marcel Goldberg

Epidemiologist, Emeritus Professor at University of Paris Descartes.

Matthieu Grall

Head of the technological expertise department of the Commission Nationale de l'Informatique et des Libertés.

Jean-Patrick Lajonchère

Managing Director of the Paris Saint-Joseph Hospital Group, chargé de mission to the Minister of Europe and Foreign Affairs for export health support.

Catherine Cheze Le Rest, M.D.,

Professor in Biophysics, Department of Nuclear Medicine, University of Poitiers.

Alain Livartowski

Oncologist, Institut Curie—Ensemble hospitalier, Deputy Director of the Data Department, member of the Institut du Thorax.

Bertrand Lukacs

Urologist, Paris.

Ozanan R. Meireles

Surgeon, MGH

Nicolas Padoy

Associate Professor on a Chair of Excellence and head of research group CAMMA (Computational Analysis and Modeling of Medical Activities) in the ICube laboratory at University of Strasbourg.

Jean-Yves Robin

Chairman of LMB Venture, Vice President of the National Federation of Third Party Trusted Numbers, former Director of ASIP Santé.

Bertrand Rondepierre

Armament engineer, now at Google.

Jacques Rouessé

Medical oncologist, honorary director of the Centre René-Huguenin de lutte contre le cancer, member of the Aca- démie nationale de Médecine.

Daniela Rus

Director of the Computer Science and Artificial Intelligence Laboratory (CSAIL), and Andrew and Erna Viterbi Professor in the Department of Electrical Engineering and Computer Science (EECS) at the Massachusetts Institute of Technology.

Elisa Salamanca

Director of the Data and Digital Innovations Departement—Information Systems Direction, AP-HP.

François Sigaux

Professor of Haematoloy, University of Paris Diderot—APHP; Executive Scientific Director of CEA's Fundamental Research.

Georges Uzbelger

IBM France, mathematician and teacher, member of the Syntec Numérique Santé.

Guy Vallancien

Member of the National Academy of Medicine, Member of the Parliamentary Office for Scientific and Technological Choices, President of CHAM.

Alain-Jacques Valleron

Epidemiologist, Professor Emeritus of Biostatistics and Medical Informatics at Sorbonne University, Member of the Academy of Sciences, Inserm Unit U1169 Genetic Therapy, Genetics, Epigenetics in Neurology, Endocrinology and Child Development.

Jean-Philippe Vert

Mathematician, Specialist in artificial intelligence and bioinformatics, research scientist at Google and adjunct professor at MINES ParisTech and École normale supérieure.

Cédric Villani

Mathematician, Fields Medal winner, professor at the Claude Bernard Lyon 1 University, member of the Academy of Sciences, Member of French Parliament, first Vice-President of the Office parlementaire d'évaluation des choix scientifiques et technologiques (OPECST).

Dimitris Visvikis

Physicist, Director of Research INSERM, LaTIM, UMR 1101, Fellow IPEM, Senior Member IEEE.

Gilles Wainrib

Scientific Director and co-founder of Owkin.

Marie Zins

Epidemiologist, University of Paris Descartes, Director of the Population Epidemiological Cohorts Unit (UMS 011 Inserm-UVSQ), in charge of the Constances National Infrastructure Biology and Health.

Footnotes

1

Villani C.,Donner un sens à l’intelligence artificielle , p. 197.

© Springer Nature Switzerland AG 2020

B. Nordlinger et al. (eds.)Healthcare and Artificial Intelligencehttps://doi.org/10.1007/978-3-030-32161-1_1

Artificial Intelligence and Tomorrow’s Health

Cédric Villani¹   and Bertrand Rondepierre² 

(1)

Assemblée Nationale, Paris, France

(2)

now at Google, Paris, France

Cédric Villani

Email: [e-mail]

The potential of artificial intelligence (AI) in the field of health is immense and has been, from the very beginning of the discipline, subject to significant work since the beginning of the 20th century. The 1970s saw an explosion in the scientific community’s interest in biomedical applications helped by the availability of distributed computing resources such as SUMEX-AIM (Stanford University Medical Experimental Computer for Artificial Intelligence in Medicine) from Stanford University and Rutgers University. Among the first practical applications to be identified was the MYCIN project. Developed in the 1960s and 1970s at Stanford University it used an expert system to identify bacteria that cause serious infections and then provided appropriate treatments.

In addition to the obvious benefits for society as a whole, the field of health has always exercised a particular fascination in the scientific community in addition to being a natural playground for artificial intelligence. The general increase in the volumes of data available in extraordinary proportions and the complexity, number, and variety of phenomena involved make health an almost infinite and extremely diversified subject of study. This complexity, which a priori defeats any attempt to fully model human biology and the mechanisms at work, is precisely the privileged place of expression for AI. Like a doctor, AI techniques are based on observations that produce information that can be used by the practitioner when confronted with theoretical and empirical knowledge. In a context where it is increasingly difficult to replicate the results of community-driven studies and where, as in many other areas, the transition from discovery to a product or new practice is rarely frictionless, AI is an exception because it is perceived as a capacity that can be directly activated and used in the field.

However, the need for theoretical knowledge is sometimes challenged by recent approaches, such as neural networks and deep learning, that manage to discover for themselves how to accomplish certain tasks without a priori information on the phenomenon under study. Based on the observation of many examples, these methods establish correlational links between the patterns observed and the ability to perform the task requested. However, such discovered links can be misleading because of the well-known difference between correlation and causality that mathematicians know well, on the one hand, and because data can carry biases that distort the relevance of the result obtained, on the other hand. Such weaknesses are not critical when it comes to recommending a movie but can have particularly disastrous consequences in the field of medicine. Autonomous discovery by algorithms is not a necessity, however, because centuries of study and deepening our knowledge of medicine have enabled us to produce a wide spectrum of models applicable to human biology unlike in some fields where human beings are perfectly incapable of explaining the cognitive processes at work to translate data into usable information. If these models exist, it must be noted that medicine as a whole remains to be based on observation and its practice is therefore extremely dependent on the experience of the practitioner whose diagnosis is based on a priori knowledge, as well as on a history of previous situations with which the new one is compared. Two practitioners in health and AI with different experiences will therefore potentially and naturally arrive at two different diagnoses from the same observation.

Like some therapeutic or epidemiological discoveries whose explanation escapes the current state of our knowledge some phenomena are only understood through an empirical process. Thus, it is at the intersection between the blind approach and full formal modeling that AI can reveal its true potential and, for example, avoid gross errors that are incomprehensible to a human being.

The importance of empirical information leads to the observation that the data that express it, even more than elsewhere, constitute the core of the practice of medicine in its relationship with the patient. It is therefore not surprising that there is an exponential increase in the quantity and diversity of available data as part of a global digitization of all medical practice and the use of ever more numerous and varied sensors. Reports written by health professionals, analytical results, genomics, medical imaging, biological signal recordings, and drug use history are examples of data sources that can be the raw material for AI research. Their diversity is an expression of the multiplicity of possible applications for health some of which can be cited as recurrent examples: diagnostic assistance, pharmacovigilance, materiovigilance, infectiovigilance, personalized medicine, patient follow-up, clinical research, and medical–administrative uses. However, the data are not in themselves an open sesame whose possession alone will allow the development of AI. Data in themselves are only of interest if they are clean, neat, and well labeled. It is for this reason that any AI approach necessarily involves a work step on these data to extract the essence of the data and thus enable them to be exploited at their true value.

The state of the art of AI for these various applications is highly heterogeneous. Like other fields, neural networks have made significant progress in the exploitation of biological signals as evidenced by the success of young French companies such as Cardiologs, which specializes in the automatic analysis of electrocardiograms. On these issues the combination of data availability and algorithm refinement has made it possible to match, if not sometimes exceed, the level of performance obtained by a human to detect pathologies from these signals. Similarly, use of such signals related to vigilance have been made possible with initial successes (particularly, in the correlation of pathologies) thanks to the depth of public databases. One example is the work published in February 2018 in The Lancet Public Health showing that excessive alcohol consumption is associated with a tripling of the risk for dementia, in general, and a doubling of the risk for developing Alzheimer disease. Advances in natural language processing have made it possible to make progress in the extraction of information (particularly, from reports written by doctors) and one hopefully to capitalize on medical information in the long term by simplifying the upstream work of qualifying and enriching data. The AI revolution for health has only